GSE264162 Bleomycin Lung PreProcessing
Kendrix Kek
2024-06-02
library(GEOquery)
library(crayon)
library(tidyverse)
library(Seurat)
library(SeuratData)
library(Azimuth)
library(patchwork)
library(Matrix)
library(RCurl)
library(DoubletFinder)
library(scales)
library(SoupX)
library(cowplot)
library(metap)
library(SingleCellExperiment)
library(DropletUtils)
library(AnnotationHub)
library(HGNChelper)
library(ensembldb)
library(multtest)
library(glmGamPoi)
library(pbapply)
library(data.table)
library(ggrepel)
library(ggpubr)
library(stringr)
library(canvasXpress)
library(clinfun)
library(GGally)
library(factoextra)
library(DESeq2)
library(limma)
library(fgsea)
library(org.Mm.eg.db)
library(kableExtra)Step 1: Data Exploration and Import
Step 1.1: Read in all sample matrix and create Seurat object
# Specify path to supplementary files.
supp_file_path = "~/Documents/Work/UnityHealth/Data_Projects/scRNAseq_LungFibrosisAnalysis/GSE264162/"
# Get unique sample names.
sample_names = list.files(path = paste0(supp_file_path, "raw_data"), pattern = "GSM", full.names = FALSE)
sample_names## [1] "GSM8213185_LG_AA_S1_1" "GSM8213185_LG_AA_S1_2" "GSM8213186_LG_AA_S2_1"
## [4] "GSM8213186_LG_AA_S2_2" "GSM8213187_LG_AA_S3_1" "GSM8213187_LG_AA_S3_2"
## [7] "GSM8213188_LG_AA_S4_1" "GSM8213188_LG_AA_S4_2" "GSM8213189_LG_AA_S5_1"
## [10] "GSM8213189_LG_AA_S5_2" "GSM8213190_LG_AA_S6_1" "GSM8213190_LG_AA_S6_2"
## [13] "GSM8213191_LG_AA_S7_1" "GSM8213191_LG_AA_S7_2"# Create Seurat object for each sample.
for (file in sample_names){
seurat_data <- Read10X(data.dir = paste0(supp_file_path, "raw_data/", file))
seurat_obj <- CreateSeuratObject(counts = seurat_data,
min.features = 100,
project = file)
assign(file, seurat_obj)
}
# Remove unneeded objects from the environment.
rm(file, seurat_data, seurat_obj)Step 1.2: Merge Seurat objects
# Merge Seurat objects.
merged_seurat <- merge(x = GSM8213185_LG_AA_S1_1,
y = c(GSM8213185_LG_AA_S1_2,
GSM8213186_LG_AA_S2_1,
GSM8213186_LG_AA_S2_2,
GSM8213187_LG_AA_S3_1,
GSM8213187_LG_AA_S3_2,
GSM8213188_LG_AA_S4_1,
GSM8213188_LG_AA_S4_2,
GSM8213189_LG_AA_S5_1,
GSM8213189_LG_AA_S5_2,
GSM8213190_LG_AA_S6_1,
GSM8213190_LG_AA_S6_2,
GSM8213191_LG_AA_S7_1,
GSM8213191_LG_AA_S7_2),
add.cell.id = c("Young_Sham_1_rep1",
"Young_Sham_1_rep2",
"Young_Sham_2_rep1",
"Young_Sham_2_rep2",
"Young_Sham_3_rep1",
"Young_Sham_3_rep2",
"Bleo_14_days_1_rep1",
"Bleo_14_days_1_rep2",
"Bleo_14_days_2_rep1",
"Bleo_14_days_2_rep2",
"Bleo_35_days_1_rep1",
"Bleo_35_days_1_rep2",
"Bleo_35_days_2_rep1",
"Bleo_35_days_2_rep2"))
# Concatenate the count matrices of the samples together.
merged_seurat <- JoinLayers(merged_seurat)
# Remove individual Seurat objects after successful merge.
rm(list = ls(pattern = "GSM"))
gc()## used (Mb) gc trigger (Mb) limit (Mb) max used (Mb)
## Ncells 15829460 845.4 25595414 1367.0 NA 25595414 1367.0
## Vcells 333236570 2542.4 1571576192 11990.2 1024000 1894557522 14454.4Step 1.3: Read metadata
# Pull GEO metadata.
GSE264162_meta <- getGEO(GEO = "GSE264162", GSEMatrix = TRUE)## Found 1 file(s)## GSE264162_series_matrix.txt.gzGSE264162_meta <- pData(GSE264162_meta$GSE264162_series_matrix.txt.gz)
# Subset metadata.
GSE264162_meta <- GSE264162_meta %>% subset(select = c(title,
geo_accession,
organism_ch1,
characteristics_ch1.5,
`age:ch1`,
`cell type:ch1`,
`genotype:ch1`,
`strain:ch1`,
`tissue:ch1`))
# Rename columns to more accessible names.
colnames(GSE264162_meta) <- c("title",
"geo_accession",
"organism",
"treatment",
"age",
"cell_type",
"genotype",
"strain",
"tissue")
# Edit the metadata columns.
GSE264162_meta$title <- gsub(" ", "_", GSE264162_meta$title)
GSE264162_meta$title <- gsub("-", "_", GSE264162_meta$title)
GSE264162_meta$treatment <- gsub("treatment: ", "", GSE264162_meta$treatment)
# Add timepoint variable.
GSE264162_meta$timepoint <- ifelse(GSE264162_meta$treatment == "Sham", "Sham", "")
GSE264162_meta$timepoint <- ifelse(GSE264162_meta$timepoint == "", paste0("Day ",
gsub("Bleo_", "",
gsub("_days_1", "",
gsub("_days_2", "",
GSE264162_meta$title)))),
GSE264162_meta$timepoint)Step 1.4: Add metadata variables from GEO
We can merged the metadata from GEO and put the merged metadata back into the Seurat object.
# Get metadata from Seurat object.
metadata = merged_seurat@meta.data %>% rownames_to_column()
dim(metadata)## [1] 86347 4# Create a variable column to merge dataframes.
metadata$geo_accession <- str_extract(metadata$orig.ident, "[^_]+")
# Merge metadata with GEO metadata by the newly created geo_accession column.
metadata <- merge(metadata, GSE264162_meta, by = "geo_accession")
dim(metadata)## [1] 86347 14# Edit and retain only required columns.
metadata <- metadata %>% column_to_rownames() %>% subset(select = -geo_accession)
# Generate quality metrics.
names(metadata)[names(metadata) == "nCount_RNA"] <- "nUMI"
names(metadata)[names(metadata) == "nFeature_RNA"] <- "nGene"
# Compute mitochondrial ratio.
metadata$mitoRatio = PercentageFeatureSet(object = merged_seurat, pattern = "^mt-")
metadata$mitoRatio = metadata$mitoRatio / 100
# Calculate novelty score.
metadata$log10GenesPerUMI <- log10(metadata$nGene) / log10(metadata$nUMI)
# Add the new metadata back to the Seurat object.
merged_seurat@meta.data <- metadata
# Remove unneeded objects.
rm(GSE264162_meta, metadata, sample_names)Step 2: Quality Control
Step 2.1: Visualize cell counts
Check cell counts with number of unique barcodes detected to determine capture efficiency.
# Extract metadata from Seurat object.
metadata = merged_seurat@meta.data
# Visualize cell counts.
metadata %>%
ggplot(aes(x = treatment, fill = treatment)) +
geom_bar() +
theme_classic() +
theme(axis.text.x = element_text(angle = 45, vjust = 1, hjust = 1)) +
theme(plot.title = element_text(hjust = 0.5, face = "bold")) +
ggtitle("NCells")
Step 2.2: Visualize UMI counts per cell
The UMI counts per cell should generally be above 500, that is the low end of what we expect. If UMI counts are between 500-1000 counts, it is usable but the cells probably should have been sequenced more deeply.
# Visualize UMI counts per cell.
metadata %>%
ggplot(aes(x = nUMI, color = treatment, fill = treatment)) +
geom_density(alpha = 0.2) +
scale_x_log10(labels = label_number()) +
ylab("Cell density") +
geom_vline(xintercept = 500, linetype = "dashed", color = "red") +
theme_classic()
Step 2.3: Visualize novelty score of cells
Generally, we expect the novelty score to be above 0.80 for good quality cells.
# Visualize novelty score.
metadata %>%
ggplot(aes(x = log10GenesPerUMI, color = treatment, fill = treatment)) +
geom_density(alpha = 0.2) +
geom_vline(xintercept = 0.8, linetype = "dashed", color = "red") +
theme_classic()
Step 2.4: Visualize mitochondrial ratio
In general, poor quality samples have mitochondrial ratio higher than 0.2 mark.
# Visualize mitochondrial ratio.
metadata %>%
ggplot(aes(x = mitoRatio, color = treatment, fill = treatment)) +
geom_density(alpha = 0.2) +
scale_x_log10() +
geom_vline(xintercept = 0.2, linetype = "dashed", color = "red") +
theme_classic()## Warning in scale_x_log10(): log-10 transformation introduced infinite values.## Warning: Removed 532 rows containing non-finite outside the scale range
## (`stat_density()`).
Step 2.5: Visualize joint filtering effects
Considering any of these QC metrics in isolation can lead to misinterpretation of cellular signals. Jointly visualizing the count and gene thresholds and additionally overlaying the mitochondrial fraction, gives a summarized persepective of the quality per cell.
# Visualize joint filtering QCs.
metadata %>%
ggplot(aes(x = nUMI, y = nGene, color = mitoRatio)) +
geom_point() +
scale_colour_gradient(low = "gray90", high = "black") +
stat_smooth(method = lm) +
scale_x_log10() +
scale_y_log10() +
geom_vline(xintercept = 500) +
geom_hline(yintercept = 250) +
theme_classic() +
facet_wrap(~treatment)## `geom_smooth()` using formula = 'y ~ x'## Warning: The following aesthetics were dropped during statistical transformation:
## colour.
## ℹ This can happen when ggplot fails to infer the correct grouping structure in
## the data.
## ℹ Did you forget to specify a `group` aesthetic or to convert a numerical
## variable into a factor?
## The following aesthetics were dropped during statistical transformation:
## colour.
## ℹ This can happen when ggplot fails to infer the correct grouping structure in
## the data.
## ℹ Did you forget to specify a `group` aesthetic or to convert a numerical
## variable into a factor?
Step 2.6: Visualize valid cells with knee plot
# Rank expression barcodes.
br.out = barcodeRanks(merged_seurat@assays$RNA$counts)
# Plot knee plot.
plot(br.out$rank, br.out$total, log = "xy", xlab = "Rank", ylab = "Total")
abline(h = metadata(br.out)$knee, col = "dodgerblue", lty = 2)
abline(h = metadata(br.out)$inflection, col = "forestgreen", lty = 2)
legend("bottomleft", lty = 2, col = c("dodgerblue", "forestgreen"), legend = c("knee", "inflection"))
# Remove unneeded object.
rm(br.out)Step 2.7: Visualize doublets
Before running DoubletFinder(), we need to run a quick
Seurat processing pipeline first.
# Run quick Seurat processing pipeline.
set.seed(10)
seur_obj <- NormalizeData(merged_seurat)
seur_obj <- FindVariableFeatures(seur_obj)
seur_obj <- ScaleData(seur_obj)
seur_obj <- RunPCA(seur_obj)
seur_obj <- FindNeighbors(seur_obj)
seur_obj <- FindClusters(seur_obj)Determine the threshold for the pANN score that corresponds to the expected number of heterotypic doublets (parameter: nEXP = expected number of heterotypic doublets).
# Calculate homotypic doublet proportion.
type.freq <- table(seur_obj@meta.data$seurat_clusters) / ncol(seur_obj)
homotypic.prop <- sum(type.freq^2)
# Calculate nEXP based on number of heterotypic doublet.
nEXP = 0.009 * (ncol(seur_obj)/1000) * (1-homotypic.prop) * ncol(seur_obj)
nEXP
# Define pN and nPC.
pN = 0.25
PCs = 25
# Determine bimodality coeffient (pK).
sweep.out <- paramSweep(seur_obj, PCs = 1:PCs, sct = FALSE)
sweep.stats <- summarizeSweep(sweep.out)Summarize sweep output to get the ideal coefficient values.
# Summarize coefficient output.
plot(sweep.stats[, 2:3])
sweep.stats[as.numeric(as.character(sweep.stats[, 2])) < 0.075 & sweep.stats[, 3] > 0.8,]
# Clear memory.
gc()
# Remove unneeded objects.
rm(homotypic.prop, sweep.out, type.freq)pK = 0.0005 yields consistently high scores with pN = 0.05 producing the highest score.
# Run DoubletFinder using coefficients from paramSweep.
pK = 0.0005
pN = 0.05
seur_obj <- doubletFinder(seur_obj, PCs = 1:PCs, pN = pN, pK = pK, nExp = nEXP)
# DoubletFinder adds the doublet score in the meta.data slot.
head(seur_obj@metadata)
# Grab the metadata and rename relevant columns.
doublet_meta = seur_obj@meta.data
names(doublet_meta)[grep("pANN", colnames(doublet_meta))] <- "pANN"
names(doublet_meta)[grep("DF.classifications", colnames(doublet_meta))] <- "doublet_class"
# Keep relevant columns only.
doublet_meta <- doublet_meta %>%
subset(select = c("orig.ident", "pANN", "doublet_class")) %>%
rownames_to_column("matched_id")
# Get metadata from the Seurat object to merge.
metadata = merged_seurat@meta.data %>%
rownames_to_column("matched_id")
# Check if matched_id for both dataframes match before merging.
identical(metadata$matched_id, doublet_meta$matched_id) # Must be TRUE.
# Merge both metadata together.
metadata <- left_join(metadata, doublet_meta, by = "matched_id") %>%
subset(select = -orig.ident.y) %>%
column_to_rownames("matched_id")
# Add merged metadata back to the Seurat object.
all(colnames(merged_seurat) == rownames(metadata)) # Must be TRUE.
merged_seurat@meta.data <- metadataStep 2.8: Filtering
Step 2.8.1: Cell-level filtering
Now that we have visualized the various metrics, we can decide on the thresholds to apply which will result in the removal of low quality cells. We will use the following thresholds:
nUMI > 500
nGene > 250
log10GenesPerUMI > 0.8
mitoRatio < 0.2
# Get pre-filtering cell counts.
dim(merged_seurat)[2] # 86,347 cells## [1] 86347# Low-quality cell filtering.
umi_filtered <- subset(x = merged_seurat, subset = nUMI >= 500)
gene_filtered <- subset(x = umi_filtered, subset = nGene >= 250)
genesperumi_filtered <- subset(x = gene_filtered, subset = log10GenesPerUMI > 0.80)
mitoratio_filtered <- subset(x = genesperumi_filtered, subset = mitoRatio < 0.20)
# Get post-filtering Seurat object.
# Enter level of filtering desired here.
cell_filtered_seurat = mitoratio_filtered
# Get counts after low-quality cell filtering.
dim(cell_filtered_seurat)[2] # 80,505 cells.## [1] 80505We can visualize the amount of cells removed and the number of cells remaining for each filtering step in the plot below.
Step 2.8.2: Gene-level filtering
We also remove genes that have zero expression in all cells and genes that are only expressed in 10 or less cells as genes that are expressed in only a handful of cells are not meaningful and can bring down the averages for all the other cells that they are not expressed in.
# Extract count matrix.
counts <- GetAssayData(object = cell_filtered_seurat, layer = "counts")
# Get a logical vector for every gene that has more than zero counts per cell.
nonzero <- counts > 0
# Get a logical vector of genes that are expressed in 10 or more cells.
keep_genes <- Matrix::rowSums(nonzero) >= 10
# Subset counts to keep genes that are non-zero and expressed in 10 or more cells.
filtered_counts <- counts[keep_genes, ]
# Create new Seurat object with the subsetted counts.
filtered_seurat <- CreateSeuratObject(filtered_counts, meta.data = cell_filtered_seurat@meta.data)We can visualize the amount of genes removed and the number of genes remaining for each filtering step in the plot below.
Step 3: Normalization
Step 3.1: Pre-normalization visualization
Let’s have a look of our expression distribution before normalization.
# Visualize expression distribution before normalization.
hist(colSums(filtered_seurat[["RNA"]]$counts),
breaks = 100,
col = "#212226",
main = "Pre-Normalization: Expression distribution",
xlab = "Sum of expression")
Step 3.2: Simple normalization
Before we make any comparisons across cells, we will apply a simple normalization. This is solely for the purpose of exploring the sources of variation in our data.
# Apply simple normalization to merged Seurat objects.
seurat_phase <- NormalizeData(filtered_seurat)## Normalizing layer: counts# Visualize expression distribution after simple normalization.
hist(colSums(seurat_phase[["RNA"]]$data),
breaks = 100,
col = "#b8520a",
main = "Post-Simple Normalization: Expression distribution",
xlab = "Sum of expression")
# Remove unneeded objects.
rm(filtered_seurat)Step 3.3: Evaluate effects of cell cycle
Step 3.3.1: Set up cell cycle scores
# Download cell cycle genes for organism at https://github.com/hbc/tinyatlas/tree/master/cell_cycle.
cc_file <- getURL("https://raw.githubusercontent.com/hbc/tinyatlas/master/cell_cycle/Mus_musculus.csv")
cell_cycle_genes <- read.csv(text = cc_file)
# Remove unneeded objects.
rm(cc_file)All of the cell cycle genes are Ensembl IDs, but our gene IDs are the gene names. To score the genes in our count matrix for cell cycle, we need to obtain the gene names for the cell cycle genes.
# Connect to AnnotationHub.
ah <- AnnotationHub()
# Access the Ensembl database for organism.
ahDb <- query(ah,
pattern = c("Mus musculus", "EnsDb"),
ignore.case = TRUE)
# Acquire the latest annotation files.
id <- ahDb %>%
mcols() %>%
rownames() %>%
tail(n = 1)
# Download the appropriate Ensembldb database.
edb <- ah[[id]]## loading from cache# Extract gene-level information from database.
annotations <- genes(edb,
return.type = "data.frame")
# Select annotations of interest.
annotations <- annotations %>%
dplyr::select(gene_id, gene_name, seq_name, gene_biotype, description)
# Remove unneeded objects.
rm(ah, ahDb, edb, id)We can merge the Ensembl genes to the cell cycle genes.
# Get gene names for Ensembl IDs for each gene.
cell_cycle_markers <- dplyr::left_join(cell_cycle_genes, annotations, by = c("geneID" = "gene_id"))
# Acquire the S phase genes.
s_genes <- cell_cycle_markers %>%
dplyr::filter(phase == "S") %>%
pull("gene_name")
# Acquire the G2M phase genes.
g2m_genes <- cell_cycle_markers %>%
dplyr::filter(phase == "G2/M") %>%
pull("gene_name")
# Remove unneeded objects.
rm(cell_cycle_genes, cell_cycle_markers, annotations)We can then use the annotated cell cycle genes to perform cell cycle scoring.
# Perform cell cycle scoring
seurat_phase <- CellCycleScoring(seurat_phase,
g2m.features = g2m_genes,
s.features = s_genes)
# Remove unneeded objects.
rm(g2m_genes, s_genes)Step 3.3.2: Use PCA to evaluate effects of cell cycle
# Identify the most variable genes.
seurat_phase <- FindVariableFeatures(seurat_phase,
selection.method = "vst",
nfeatures = 2000,
verbose = FALSE)
# Scale the counts.
seurat_phase <- ScaleData(seurat_phase)## Centering and scaling data matrix# Visualize expression distribution after scaling.
hist(colSums(seurat_phase[["RNA"]]$scale.data),
breaks = 100,
col = "#779e7f",
main = "Post-Scaling: Expression distribution",
xlab = "Sum of expression")
# Identify the 15 most highly variable genes.
ranked_variable_genes <- VariableFeatures(seurat_phase)
top_genes <- ranked_variable_genes[1:15]
# Plot the average expression and variance of these genes with labels to indicate which genes are in the top 15.
p <- VariableFeaturePlot(seurat_phase)
LabelPoints(plot = p, points = top_genes, repel = TRUE)## When using repel, set xnudge and ynudge to 0 for optimal results
# Remove unneeded objects.
rm(ranked_variable_genes, top_genes, p)# Perform PCA.
seurat_phase <- RunPCA(seurat_phase)## PC_ 1
## Positive: Ccdc153, Tmem212, 1110017D15Rik, Fam183b, Rsph1, Dynlrb2, Foxj1, Cfap126, Cyp2s1, Arhgdig
## Sec14l3, 1700007K13Rik, Cd24a, Ccdc113, Mlf1, Gm19935, Riiad1, 1700016K19Rik, Aldh3b1, 1700094D03Rik
## Pifo, Tspan1, BC051019, Cfap53, Cdhr3, Mns1, Tctex1d4, Sntn, Ak7, Traf3ip1
## Negative: Sparc, Vim, Igfbp7, Col4a1, Bgn, Mgp, Apoe, Col6a1, Adamts1, Col3a1
## Col1a2, Fstl1, Sparcl1, Col1a1, Lgals1, Eln, Ltbp4, Igfbp4, Fbn1, Serpinh1
## Col5a2, Col4a2, Selenom, Cfh, Loxl1, Serpina3n, Ptgis, Lrp1, Emp3, Colec12
## PC_ 2
## Positive: Sftpa1, Napsa, Cxcl15, Sftpd, Slc34a2, Lyz2, Lamp3, Cd74, Scd1, S100g
## Ager, Cldn18, Fabp5, H2-Aa, Chil1, Hc, Egfl6, Aqp5, Elovl1, Ctsh
## H2-Ab1, Bex4, Lrp2, Acoxl, Slc39a8, Chia1, Mal, Lcn2, Wfdc2, Bex2
## Negative: Col3a1, Col1a2, Serping1, Col1a1, Bgn, Mgp, Gsn, Loxl1, Lrp1, Lgals1
## Fbn1, Rbp1, Plxdc2, Mfap5, Mfap4, Pmepa1, Col5a2, Fn1, Mmp2, Fstl1
## Adamts2, Rarres2, Olfml3, Sod3, Sparcl1, Dpt, Col6a3, Colec12, Ptgis, Cfh
## PC_ 3
## Positive: Sftpd, Napsa, Cxcl15, Chil1, Slc34a2, Scd1, Sftpa1, Lamp3, Lyz2, S100g
## Cldn18, Ager, Egfl6, Hc, Fabp5, Aqp5, Selenbp1, Mgst1, Elovl1, Ctsh
## Bex4, Wfdc2, Lrp2, Hp, Cd74, Col6a1, Acoxl, Prdx6, Apoc1, Il33
## Negative: Ramp2, Cdh5, Cldn5, Egfl7, Calcrl, Pecam1, Ctla2a, Ptprb, Aqp1, Ehd4
## Rasip1, Ecscr, Flt1, Eng, Icam2, Plvap, S1pr1, Srgn, Cd93, Ly6a
## Tspan7, Cyyr1, Podxl, Lyve1, Cavin2, Slfn5, Gimap6, Acvrl1, Cemip2, Clec14a
## PC_ 4
## Positive: Inmt, Scn7a, Fmo2, Ogn, Ly6a, Tmem100, Pcolce2, Clec3b, Fmo1, Hpgd
## Dpep1, Plpp3, Mfap4, Acvrl1, C7, Mamdc2, Clec14a, Tspan7, Flt1, Sept4
## Tcf21, Atp1a2, Apoe, Ace, Iigp1, Gsn, Thbd, Hsd11b1, Gstm1, Abca8a
## Negative: Birc5, Mki67, Tpx2, Cdca8, Cdk1, Pclaf, Ccna2, Ube2c, Cdca3, Hmmr
## Top2a, Cenpf, Racgap1, Prc1, Cenpe, Anln, Ccnb1, Cenpa, Pbk, Plk1
## Ccnb2, Cdc20, Aurkb, Nusap1, Ckap2l, Knl1, Kif11, Knstrn, Cdkn3, Tacc3
## PC_ 5
## Positive: Fabp5, Chil1, Lyz2, Slc34a2, Lamp3, Cxcl15, Sftpa1, S100g, Hc, Scd1
## Cbr2, Acoxl, Lcn2, Il33, Egfl6, Lrg1, Bex2, H2-Aa, Napsa, Apoc1
## Elovl1, Itih4, Lrp2, Hp, Sftpd, Chia1, Bex4, H2-Ab1, Trf, Aox3
## Negative: Rtkn2, Col4a3, Akap5, Spock2, Scnn1g, Col4a4, Igfbp2, Krt7, Itgb6, Hopx
## Ndnf, Abca5, Flrt3, Clic5, Tmem37, Ano1, 2210011C24Rik, Gprc5a, Lama3, Cdkn2b
## Cryab, Emp2, Lamc2, Hck, Pxdc1, Pdpn, Nt5e, Tspan8, Fam189a2, Scnn1b# Plot the PCA colored by cell cycle phase.
DimPlot(seurat_phase,
reduction = "pca",
group.by = "Phase",
split.by = "Phase")
# Split effects of PCA by treatment.
DimPlot(seurat_phase,
reduction = "pca",
group.by = "Phase",
split.by = "timepoint")
We don’t see large differences due to the effects of cell cycle, so cell cycle effects are not regressed out.
Step 3.4: SCTransform normalization
# Free up memory.
gc()## used (Mb) gc trigger (Mb) limit (Mb) max used (Mb)
## Ncells 16322719 871.8 25595414 1367.0 NA 25595414 1367.0
## Vcells 776961035 5927.8 1515293976 11560.8 1024000 1894557522 14454.4# Set memory limit.
options(future.globals.maxSize = 8000 * 1024^2)
# Perform SCTransform normalization.
seurat_sct <- SCTransform(seurat_phase,
vars.to.regress = c("mitoRatio"),
vst.flavor = "v2",
conserve.memory = TRUE,
verbose = FALSE)## Will not return corrected UMI because residual type is not set to 'pearson'# Check which objects are stored in which assays.
seurat_sct@assays## $RNA
## Assay (v5) data with 22546 features for 80505 cells
## Top 10 variable features:
## Bpifa1, Scgb3a1, Bpifb1, Ltf, Reg3g, Cxcl13, Spp1, Retnla, Tff2, Saa3
## Layers:
## counts, data, scale.data
##
## $SCT
## SCTAssay data with 22546 features for 80505 cells, and 1 SCTModel(s)
## Top 10 variable features:
## Sftpc, Scgb1a1, Lyz1, Spp1, Mgp, Ccl21a, Scgb3a2, Lyz2, Bpifa1, Scgb3a1# Remove unneeded objects.
rm(seurat_phase)The SCTransform normalized counts are stored in the SCT$counts slot. Here is the distribution of the SCTransform normalized counts distribution.
# Visualize expression distribution after scaling.
hist(colSums(seurat_sct[["SCT"]]$counts),
breaks = 100,
col = "#332f42",
main = "Post-SCTransform: Expression distribution",
xlab = "Sum of expression")
The log-normalized SCTransfromed counts are also stored in the SCT$data slot. Here is the distribution of the log-normalized SCTransformed counts.
hist(colSums(seurat_sct[["SCT"]]$data),
breaks = 100,
col = "#b14b34",
main = "Post-SCTransformed: Log-normalized expression distribution",
xlab = "Sum of expression")
Step 4: Dimensional Reduction and Clustering
The goal of this step is to perform unsupervised clustering to identify groups of cells based on similarities of their transcriptomes without any prior knowledge of the labels or the number of clusters. This can be challenging due to the high level of noise and large number of dimensions, so it is often beneficial to apply a dimensionality reduction method to reduce the amount of noise.
Step 4.1: Run PCA
# Run PCA.
seurat_sct <- RunPCA(seurat_sct, verbose = FALSE)
# Visualize dimension reduction from PCA.
seurat_sct <- RunUMAP(seurat_sct, dims = 1:40, reduction = "pca")## Warning: The default method for RunUMAP has changed from calling Python UMAP via reticulate to the R-native UWOT using the cosine metric
## To use Python UMAP via reticulate, set umap.method to 'umap-learn' and metric to 'correlation'
## This message will be shown once per session## 11:38:05 UMAP embedding parameters a = 0.9922 b = 1.112## 11:38:05 Read 80505 rows and found 40 numeric columns## 11:38:05 Using Annoy for neighbor search, n_neighbors = 30## 11:38:05 Building Annoy index with metric = cosine, n_trees = 50## 0% 10 20 30 40 50 60 70 80 90 100%## [----|----|----|----|----|----|----|----|----|----|## **************************************************|
## 11:38:12 Writing NN index file to temp file /var/folders/1c/sbnd3yvd7mzg5lng1jcr70tw0000gn/T//RtmpG4X7bv/fileaad2c52b519
## 11:38:12 Searching Annoy index using 1 thread, search_k = 3000
## 11:38:38 Annoy recall = 100%
## 11:38:40 Commencing smooth kNN distance calibration using 1 thread with target n_neighbors = 30
## 11:38:43 Initializing from normalized Laplacian + noise (using RSpectra)
## 11:38:48 Commencing optimization for 200 epochs, with 3408926 positive edges
## 11:39:21 Optimization finishedDimPlot(object = seurat_sct, group.by = "treatment")
Step 4.2: Top PC-driving metagenes
# Explore the top PC markers.
DimHeatmap(seurat_sct,
dims = 1:9,
cells = 500,
balanced = TRUE)
We can visualize an elbow plot to see how many PCs to use for clustering.
# Build the elbow plot.
# Determine percent of variation associated with each PC.
pct <- seurat_sct[["pca"]]@stdev / sum(seurat_sct[["pca"]]@stdev) * 100
# Calculate cumulative percents for each PC.
cumu <- cumsum(pct)
# Determine which PC exhibits cumulative percent greater than 90% and % variation associated with the PC as less than 5.
co1 <- which(cumu > 90 & pct < 5)[1]
co1## [1] 41# Determine the difference between variation of PC and subsequent PC.
co2 <- sort(which((pct[1:length(pct) - 1] - pct[2:length(pct)]) > 0.1), decreasing = T)[1] + 1
# Last point where change of % of variation is more than 0.1%.
co2## [1] 25# Minimum of the two calculation.
pcs <- min(co1, co2)
pcs## [1] 25# Create a dataframe with values.
elbowplot_df <- data.frame(pct = pct,
cumu = cumu,
rank = 1:length(pct))
# Elbow plot to visualize.
ggplot(elbowplot_df, aes(cumu, pct, label = rank, color = rank > pcs)) +
geom_text() +
geom_vline(xintercept = 90, color = "grey") +
geom_hline(yintercept = min(pct[pct > 5]), color = "grey") +
theme_bw()
# Remove unneeded objects.
rm(co1, co2, cumu, elbowplot_df, pcs, pct)Based on the elbow plot, we will use approx. 15 PCs to generate the clusters.
Step 4.3: Cluster the cells
The basic steps towards complete clustering involve:
Measure of similarity: How does one quantify how close two data points are?
Quality function: How does one define the clustering/partition of the points?
Algorithm: How to find the clustering whereby the quality function is optimized?
Step 4.3.1: Find neighbours
We use a graph-based community detection method where each vertex represents a cell and the weight of the edge measures similarities between two cells. Put simply, we start by building an unweighted K-nearest neighbour (k-NN) graph and add weights to obtain a shared nearest neighbour (SNN) graph. Weight can be added either by finding out the shared nodes (overlap) between two neighbours (more nodes; more similar the neighbours; more weight) or measuring the closeness of both neighbours to each other.
# Compute SNN graph.
seurat_cluster <- FindNeighbors(object = seurat_sct, dims = 1:15) # Based on elbow plot.## Computing nearest neighbor graph## Computing SNN# Remove unneeded object.
rm(seurat_sct)Step 4.3.2: Find clusters
Next, we want to iteratively group the neighbours together with the goal of combining the cells with similar transcriptomes together. Without knowing the number of clusters a priori, we use a range of resolution based on the approximation of the recommendation where datasets of 3,000 - 5,000 cells should set the resolution between 0.4 - 1.4. We have approximately 90,000 cells, so our range will start from 1.5 - 2.5.
# Determine start and end resolution range.
start_reso_range = 0.2
end_reso_range = 1.0
range_interval = 0.2
# Find clusters given the range.
seurat_cluster <- FindClusters(object = seurat_cluster,
resolution = seq(from = start_reso_range,
to = end_reso_range,
by = range_interval))## Modularity Optimizer version 1.3.0 by Ludo Waltman and Nees Jan van Eck
##
## Number of nodes: 80505
## Number of edges: 2500763
##
## Running Louvain algorithm...
## Maximum modularity in 10 random starts: 0.9649
## Number of communities: 13
## Elapsed time: 34 seconds
## Modularity Optimizer version 1.3.0 by Ludo Waltman and Nees Jan van Eck
##
## Number of nodes: 80505
## Number of edges: 2500763
##
## Running Louvain algorithm...
## Maximum modularity in 10 random starts: 0.9498
## Number of communities: 23
## Elapsed time: 27 seconds
## Modularity Optimizer version 1.3.0 by Ludo Waltman and Nees Jan van Eck
##
## Number of nodes: 80505
## Number of edges: 2500763
##
## Running Louvain algorithm...
## Maximum modularity in 10 random starts: 0.9379
## Number of communities: 26
## Elapsed time: 31 seconds
## Modularity Optimizer version 1.3.0 by Ludo Waltman and Nees Jan van Eck
##
## Number of nodes: 80505
## Number of edges: 2500763
##
## Running Louvain algorithm...
## Maximum modularity in 10 random starts: 0.9275
## Number of communities: 32
## Elapsed time: 32 seconds
## Modularity Optimizer version 1.3.0 by Ludo Waltman and Nees Jan van Eck
##
## Number of nodes: 80505
## Number of edges: 2500763
##
## Running Louvain algorithm...
## Maximum modularity in 10 random starts: 0.9193
## Number of communities: 39
## Elapsed time: 28 seconds# Remove unneeded objects.
rm(start_reso_range, end_reso_range, range_interval)We can look at the maximum number of clusters found at each resolution with the output below.
# Create a dummy table to contain the max cluster for each resolution.
maxclust = c()
clustoutput = seurat_cluster@meta.data %>% subset(select = grep("SCT_snn_", colnames(.)))
clustoutput[] <- lapply(clustoutput, function(x) as.numeric(as.character(x)))
# Iterate and print the maximum cluster count for each resolution.
for (i in 1:ncol(clustoutput)){
maxclust[i] = max(clustoutput[, i])
print(paste0("Max n-cluster for ",
names(clustoutput[i]), ": ",
maxclust[i]))
}## [1] "Max n-cluster for SCT_snn_res.0.2: 12"
## [1] "Max n-cluster for SCT_snn_res.0.4: 22"
## [1] "Max n-cluster for SCT_snn_res.0.6: 25"
## [1] "Max n-cluster for SCT_snn_res.0.8: 31"
## [1] "Max n-cluster for SCT_snn_res.1: 38"# Remove unneeded objects.
rm(clustoutput, i, maxclust)We can visualize the difference clusterings of the resolutions side by side. In this case, we will look at these resolutions: 0.6, 0.8, 1.0.
# Select resolutions to visualize.
resolutions = c("0.6", "0.8", "1")
feature_cols = paste0("SCT_snn_res.", resolutions)
# Visualize UMAP of different resolutions.
for (reso in feature_cols) {
Idents(seurat_cluster) <- reso
p <- DimPlot(seurat_cluster,
reduction = "umap",
label = TRUE,
label.size = 3) +
ggtitle(reso) +
NoLegend()
print(p)
}
# Remove unneeded objects.
rm(feature_cols, p, reso, resolutions)Step 4.4: Explore clustering output
Let’s start with the exploration with the highest resolution in the range, 1.0.
# Sort the cluster labels in ascending order.
reso_columns = grep("SCT_snn_", colnames(seurat_cluster@meta.data), value = TRUE)
metadata <- seurat_cluster@meta.data
for (col in colnames(metadata)) {
if (col %in% reso_columns){
metadata[, col] <- factor(metadata[, col], levels = paste(sort(as.integer(levels(metadata[, col])))))
}
}
seurat_cluster@meta.data <- metadata
# Assign desired cluster resolution as identity of Seurat.
Idents(object = seurat_cluster) <- "SCT_snn_res.1"
# Visualize cluster with a UMAP.
DimPlot(seurat_cluster,
reduction = "umap",
label = TRUE,
label.size = 3,
repel = TRUE) +
ggtitle("SCT_snn_res.1.0") +
NoLegend()
# Remove unneeded objects.
rm(col, metadata, reso_columns)Step 4.5: Explore clustering quality control
Step 4.5.1: Distribution of cells by treatment
# Extract number of cells per cluster.
n_cells <- FetchData(seurat_cluster,
vars = c("ident", "treatment")) %>%
dplyr::count(ident, treatment)
# Visualize distribution of cells by cluster in a barplot.
ggplot(n_cells, aes(x = ident, y = n, fill = treatment)) +
geom_bar(position = position_dodge(),
stat = "identity") +
geom_text(aes(label = n),
size = 2,
vjust = -0.2,
position = position_dodge(1)) +
theme_classic() +
theme(axis.text.x = element_text(size = 5))
# Visualize distribution of treatment in each cluster.
DimPlot(seurat_cluster,
label = TRUE,
split.by = "treatment",
label.size = 3,
repel = TRUE) +
theme(axis.text.x = element_text(size = 6),
axis.text.y = element_text(size = 6)) +
NoLegend()
# Remove unneeded object.
rm(n_cells)Step 4.5.2: Mitochondrial enrichment in clusters
We would like to see if there is any enrichment of mitochondria in our clusters and if we need to regress out the mitochondrial level in our data.
# Plot mitoRatio to see if any cluster is enriched in mitochondrial level.
FeaturePlot(seurat_cluster,
reduction = "umap",
features = "mitoRatio",
cols = c("#d6cfc7", "#01013f"),
pt.size = 0.5,
order = TRUE,
min.cutoff = "q10",
label = TRUE,
label.size = 3,
label.color = "#960000") 
There are certain clusters enriched with mitochondria, but they are within the acceptable ratio of under 0.20.
Step 4.5.3: Variation in clusters due to cell cycle phase
If our cell clusters show large differences in cell cycle expression, we might want to regress out the effects of cell cycle phase.
# Explore whether clusters segregate by cell cycle phase.
DimPlot(seurat_cluster,
label = TRUE,
split.by = "Phase",
label.size = 4,
repel = TRUE) +
theme(axis.text.x = element_text(size = 6),
axis.text.y = element_text(size = 6)) +
NoLegend()
We don’t see any noticeable cell cycle phase effects in our
clusters.
Step 5: Cell Typing
Step 5.1: Exploration of cluster identities
Step 5.1.1: Explore PC metagenes
We can explore if any specific PCs drive the separation of the clusters.
# Define columns in the metadata to visualize.
columns = c(paste0("PC_", 1:12), "ident", "umap_1", "umap_2")
# Extract data from Seurat object based on column names.
pc_data <- FetchData(seurat_cluster, vars = columns)
# Add cluster label to the centre of cluster on UMAP.
umap_label = FetchData(seurat_cluster,
vars = c("ident", "umap_1", "umap_2")) %>%
group_by(ident) %>%
summarise(x = mean(umap_1), y = mean(umap_2))
# Plot a UMAP plot for each of the PCs.
map(paste0("PC_", 1:12), function(pc){
ggplot(pc_data, aes(umap_1, umap_2)) +
geom_point(aes_string(color = pc), alpha = 0.7) +
scale_color_gradient(guide = "none", low = "#d6cfc7", high = "#01013f") +
geom_text(data = umap_label, aes(label = ident, x, y), color = "#960000", size = 4) +
theme_classic() +
ggtitle(pc)}) %>%
plot_grid(plotlist = .)## Warning: `aes_string()` was deprecated in ggplot2 3.0.0.
## ℹ Please use tidy evaluation idioms with `aes()`.
## ℹ See also `vignette("ggplot2-in-packages")` for more information.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.
# Remove unneeded objects.
rm(columns, pc_data, umap_label)We can print the genes driving different PCs and correlate those metagenes with the identity of the clusters.
# Print PC metagenes.
print(seurat_cluster[["pca"]], dims = 1:12, nfeatures = 5)## PC_ 1
## Positive: Mgp, Col3a1, Gsn, Col1a2, Col1a1
## Negative: Sftpc, Lyz2, Sftpa1, Cxcl15, Sftpb
## PC_ 2
## Positive: Mgp, Sftpc, Col3a1, Lyz2, Col1a2
## Negative: Sec14l3, Cyp2f2, AU040972, Tmem212, Aldh1a1
## PC_ 3
## Positive: Ctla2a, Vwf, Cldn5, Lyve1, Ptprb
## Negative: Mgp, Aldh1a1, Cyp2f2, Sec14l3, AU040972
## PC_ 4
## Positive: Ager, Hopx, Rtkn2, Igfbp2, S100a6
## Negative: Gsn, Dcn, Vwf, Lyz2, Sftpc
## PC_ 5
## Positive: Gsn, Inmt, Ager, Hopx, Mfap4
## Negative: Spp1, Timp1, Ccl21a, Acta2, Fn1
## PC_ 6
## Positive: Spp1, Vwf, Timp1, Acta2, Tagln
## Negative: Ccl21a, Mmrn1, Prox1, Flt4, Reln
## PC_ 7
## Positive: Dcn, Col1a2, Igfbp4, Col1a1, Pi16
## Negative: Inmt, Mfap4, Scgb1a1, Gpx3, Npnt
## PC_ 8
## Positive: Inmt, Ager, Sftpc, Hopx, Sec14l3
## Negative: Scgb1a1, Scgb3a2, Scgb3a1, Bpifa1, Cyp2f2
## PC_ 9
## Positive: Car4, Cldn5, Tmem100, Kdr, Dcn
## Negative: Vwf, Selp, Fabp4, Vcam1, Aqp1
## PC_ 10
## Positive: Acta2, Tagln, Gsn, Tpm2, Dcn
## Negative: Serpina3n, Eln, Mgp, Saa3, Fn1
## PC_ 11
## Positive: Lyz1, Cbr2, Malat1, Thbs1, Cdkn1c
## Negative: S100a6, Cd177, Wfdc2, AU040972, Cxcl17
## PC_ 12
## Positive: Lyz2, Eln, Col1a1, Scgb1a1, Col1a2
## Negative: Lyz1, S100a6, Cd177, Clu, Wfdc2Step 5.2: Explore cell identities with canonical markers
Step 5.2.1: Explore AT1 clusters
The SCTransformed data was performed only on the 3000 most variable genes. Ideally, we want to explore the identity of the clusters with all the genes in the dataset and not just the top 3000 most variable genes, so we need to set the default assay back to the RNA assay.
# Set default assay back to RNA.
DefaultAssay(seurat_cluster) <- "RNA"We can explore canonical cell type markers of our cell types of interest and visualize their expression on our UMAP. Canonical markers are obtained from LungMAP and A molecular cell atlas of the human lung from single-cell RNA sequencing (2020) Nature
AT1 cells are defined by the cell markers: Hopx, Ager, Rtkn2, Pdpn and Clic5. We can visualize both the UMAP and Dotplot to see which cluster(s) do these markers localize.
# Visualize AT1 markers expression on dotplot.
DotPlot(seurat_cluster,
features = c("Hopx", "Ager", "Rtkn2", "Pdpn", "Clic5"),
cluster.idents = TRUE,
cols = c("#d6cfc7", "#01013f"),
dot.scale = 6,
scale.by = "radius") +
ggtitle("AT1 markers") +
theme(plot.title = element_text(hjust = 0.5))
